EP4012620A1 - Method for automatically learning by transfer - Google Patents

Abstract

A computer-implemented method of machine transfer learning comprising the steps of:- training (401) an artificial neural network (R_S) to solve a first classification source problem, from at least a first set of source training data (DA_S), - updating (404) the synaptic weights of the last layer of the trained artificial neural network, from 'a second set of target training data (DA_c), for solving a second target classification problem,- For each neuron of the last layer of the artificial neural network associated with a class, said synaptic weights being proportional to the normalized centered mean of the representations of the second set of target training data (DA_c) extracted from the last hidden layer of the artificial neural network and belonging to said class.

Description

The invention relates to the field of automatic learning methods and in particular that of artificial neural networks.

The invention relates more specifically to the field of automatic transfer learning, which aims to transfer knowledge from a source task to a target task. The general issue of transfer learning can be seen as the ability of a system to recognize and apply knowledge, learned from previous tasks, to new tasks or domains that share similarities.

In the case of artificial neural networks, transfer learning consists, for example, in learning to solve a particular problem from a network that has previously learned to solve a more general problem.

In the field of automatic learning by transfer applied to artificial neural networks, a problem to be solved consists in optimizing a pre-learned network to solve a general problem in order to adapt it to a more specific problem and this by limiting the resources. computational both in terms of computation time allocated to learning the particular problem and memory resources to store the data used to perform the learning.

In other words, an objective sought in this field is to be able to carry out automatic learning for a particular problem by benefiting from the learning of a network for a more general problem and thus by limiting the complexity of implementing the learning of the particular problem.

Another problem lies in the availability of large amounts of training data which are generally required to perform efficient machine learning.

For some applications, training data can be expensive to acquire and is not always available in large quantities. Thus, there is a need to retrain the use of pre-learned neural networks on general training databases, with a minimum of adaptation, to use them to solve a specific problem from new training data in limited quantity.

This need is all the more present as many pre-trained network models are now available online. A pre-trained network is a neural network whose parameters (network weights or synaptic coefficients, in particular) have been determined after learning on a predefined learning database and to solve a given problem. .

In particular, it may be a classification problem. For example, a network pre-trained to recognize cars in a scene can be used later to solve a more specific problem of classifying the types of cars identified.

In this context, another problem faced by the user is to choose the pre-trained network which is the most suitable for solving the specific problem when several networks and/or several training databases are available.

A similar problem is to select the training data that best adapts the performance of the pre-trained network to solve the initial problem.

In general, the optimization of the pre-trained network with the aim of solving a specific problem must be carried out in such a way as to limit the number of operations and the memory space occupied.

In the field of learning transfer, the known solutions generally consist in carrying out classical learning, by means of a back-propagation algorithm based on a gradient method, for at least the last hidden layer of the neural network pre -learned by adapting the last layer of the network to the specific new problem to be solved.

This method has the disadvantage of being expensive in operations and in memory space because it requires carrying out a complete learning of at least the last hidden layer and this for all the pre-learned networks and/or all the new data of learning available.

The invention aims to solve the various aforementioned problems by proposing an optimization criterion which aims to determine the parameters of the last hidden layer of a pre-learned network for a source problem, from new learning data used to solve a target problem.

For this, the invention consists in replacing a conventional learning of the network by an update of the synaptic weights of the last layer from the new learning data and from the target problem. The invention applies advantageously to classification problems.

It thus makes it possible to limit the cost of the operations of the second learning by replacing it with a direct optimization of the parameters of the last layer of the network.

The invention thus allows the selection, with a reduced processing cost, of the pre-learned network and/or of the learning data making it possible to solve the target problem with the best performance.

• Entrainer un réseau de neurones artificiel à résoudre un premier problème source de classification, à partir d'au moins un premier ensemble de données d'apprentissage source,
• Mettre à jour les poids synaptiques de la dernière couche du réseau de neurones artificiel entrainé, à partir d'un second ensemble de données d'apprentissage cible, pour résoudre un second problème cible de classification,
• Pour chaque neurone de la dernière couche du réseau de neurones artificiel associé à une classe, lesdits poids synaptiques étant proportionnels à la moyenne centrée normalisée des représentations du second ensemble de données d'apprentissage cible extraites de la dernière couche cachée du réseau de neurones artificiel et appartenant à ladite classe.

The subject of the invention is a method, implemented by computer, of automatic learning by transfer comprising the steps of:

• Training an artificial neural network to solve a first source classification problem, from at least a first set of source training data,
• Update the synaptic weights of the last layer of the trained artificial neural network, from a second target training dataset, to solve a second target classification problem,
• For each neuron of the last layer of the artificial neural network associated with a class, said synaptic weights being proportional to the normalized centered mean of the representations of the second set of target learning data extracted from the last hidden layer of the artificial neural network and belonging to that class.

According to a particular aspect, the method according to the invention further comprises a step of replacing the last layer of the artificial neural network trained by a layer specific to the second target classification problem.

According to a particular aspect of the invention, the specific layer is a layer based on the softmax function.

According to a particular aspect of the invention, to update the synaptic weights of the last layer of the trained artificial neural network, the target learning data of the second set are propagated from the input of said network to its last layer hidden so as to be able to extract said representations.

According to a particular aspect of the invention, the artificial neural network is trained for several sets of source learning data, in order to produce several trained neural networks, the update of the synaptic weights is carried out for each trained neural network and the method further comprises a step of selecting the trained neural network which exhibits the best performance for solving the second target classification problem.

According to a particular aspect of the invention, the second set of target training data (is of smaller size than a set of source training data.

According to a particular aspect of the invention, the updating of the synaptic weights is carried out for several sets of target learning data, the second target problem is identical to the first source problem and the method further comprises a step of selecting the the target training dataset that achieves the worst performance in solving the second target classification problem.

In a variant embodiment, the method according to the invention further comprises a new step of training the artificial neural network from the set of target training data selected or from a combination of the set of data selected target training dataset and the source training dataset.

According to a particular aspect of the invention, the last hidden layer of the network of artificial neurons comprises an odd activation function, for example a hyperbolic tangent function.

The invention also relates to a computing device for implementing an artificial neural network having learned to solve a target classification problem using a learning transfer method according to the invention.

The invention also relates to a computer program comprising instructions for the execution of the method according to the invention, when the program is executed by a processor as well as a recording medium readable by a processor on which is recorded a program comprising instructions for the execution of the method according to the invention, when the program is executed by a processor.

• [Fig. 1] la figure 1 représente un schéma général d'un exemple de réseau de neurones artificiel configuré pour résoudre un problème général de classification dit problème source,
• [Fig. 2] la figure 2 représente un schéma du réseau de neurones artificiel de la figure 1 modifié pour résoudre un problème spécifique de classification, dit problème cible,
• [Fig. 3] la figure 3 représente les deux dernières couches du réseau de la figure 2,
• [Fig. 4] la figure 4 illustre, sur un organigramme, les étapes de mise en œuvre d'une méthode d'apprentissage par transfert selon un premier mode de réalisation de l'invention,
• [Fig. 5] la figure 5 illustre, sur un organigramme, les étapes de mise en œuvre d'une méthode d'apprentissage par transfert selon un deuxième mode de réalisation de l'invention,
• [Fig. 6] la figure 6 illustre, sur un organigramme, les étapes de mise en œuvre d'une méthode d'apprentissage par transfert selon un troisième mode de réalisation de l'invention,
• [Fig. 7] la figure 7 illustre un exemple comparatif de résultats obtenus avec une méthode de l'art antérieur et avec la méthode selon l'invention.

Other characteristics and advantages of the present invention will appear better on reading the following description in relation to the following appended drawings.

• [ Fig. 1 ] the figure 1 represents a general diagram of an example of an artificial neural network configured to solve a general classification problem called the source problem,
• [ Fig. 2 ] the figure 2 represents a diagram of the artificial neural network of the figure 1 modified to solve a specific classification problem, known as the target problem,
• [ Fig. 3 ] the picture 3 represents the last two layers of the network of the picture 2 ,
• [ Fig. 4 ] the figure 4 illustrates, on a flowchart, the steps for implementing a transfer learning method according to a first embodiment of the invention,
• [ Fig. 5 ] the figure 5 illustrates, on a flowchart, the steps for implementing a transfer learning method according to a second embodiment of the invention,
• [ Fig. 6 ] the figure 6 illustrates, on a flowchart, the steps for implementing a transfer learning method according to a third embodiment of the invention,
• [ Fig. 7 ] the figure 7 illustrates a comparative example of results obtained with a method of the prior art and with the method according to the invention.

The figure 1 represents a general diagram of an artificial neural network, for example a convolutional neural network. A neural network is conventionally composed of several layers C _e ,C _l ,C _l+i ,C _L ,C _s of interconnected neurons. The network comprises at least one input layer C _e and an output layer C _s and several intermediate layers C _l , C _{l +i} , C _L , also called hidden layers. The neurons N _i,e of the input layer C _e receive input data. The input data can be of different natures depending on the targeted application. For example, the input data is pixels of an image. A neural network has the general function of learning to solve a given problem, for example a problem of classification, regression, recognition, identification. A neural network is, for example, used in the field of image classification or image recognition or more generally the recognition of characteristics which may be visual, audio, textual.

Each neuron of a layer is connected, by its input and/or its output, to all the neurons of the preceding or following layer. More generally, a neuron can only be connected to part of the neurons of another layer, in particular in the case of a convolutional network. The connections between two neurons N _i,e , N _i,l of two successive layers are made through artificial synapses S ₁ , S ₂ , S ₃ which can be produced, in particular, by digital memories or by memristive devices. Synapse coefficients are optimized through a neural network learning mechanism. The learning mechanism aims to train the neural network to solve a defined problem. This mechanism comprises two distinct phases, a first phase of propagation of data from the input layer to the output layer and a second phase of back-propagation of errors from the output layer to the input layer with, for each layer, an update of the synapse weights. The errors calculated by the output neurons at the end of the data propagation phase are linked to the problem to be solved. It is, in a way general, to determine an error between the value of an output neuron and an expected value or a target value depending on the problem to be solved.

During the first phase of data propagation, learning data, for example reference images or sequences of images, are supplied as input to the neurons of the input layer and propagated in the network. Each neuron implements, during this first phase, a function of integrating the data received which consists, in the case of a convolutional network, of calculating a sum of the data received weighted by the coefficients of the synapses. In other words, each neuron performs a convolution operation on a convolution filter size corresponding to the size of the submatrix of neurons of the previous layer to which it is connected. Each neuron then propagates the result of the convolution to the neurons in the next layer. Depending on the models of neurons chosen, the integration function that it performs may vary.

The neurons N _i,s of the output layer C _s perform additional processing in that they calculate an error between the output value of the neuron N _i,s and an expected value or a target value which corresponds to the state end of the neuron of the output layer that one wishes to obtain in relation to the learning input data and the problem to be solved by the network. For example, if the network has to solve a classification problem, the expected end state of a neuron corresponds to the class it is supposed to identify in the input data.

More generally, each particular problem to be solved (classification, regression, prediction) is associated with one or more cost function(s) or objective function to be optimized (by minimization or maximization). One objective of the error back-propagation phase is to search for network parameters that optimize (eg minimize) the objective function.

The objective function depends on the data propagated between each layer of the network from the input layer but also on other parameters, for example the identifiers of the classes to which the data belong in the case of a classification problem.

During the second error back-propagation phase, the neurons of the output layer C _s transmit the calculated errors to the neurons of the previous layer C _l+1 which calculate a local error from the back-propagated error of the previous layer and in turn transmit this local error to the previous layer C _l . In parallel, each neuron calculates, from the local error, an update value of the weights of the synapses to which it is connected and updates the synapses. The process continues for each layer of neurons up to the penultimate layer which is responsible for updating the weights of the synapses which connect it to the input layer C _e .

The figure 1 represents an example of a neural network given by way of non-limiting illustration. In general, each intermediate layer of a network can be totally or partially connected to another layer.

Network input data can be organized in the form of matrices. The network can be configured to output a score or classification information or any other type of output suitable for the problem to be solved.

For example, if the input data are images or features extracted from images or image sequences, a neural network can be configured to recognize the objects extracted from the images and classify them according to different predefined classes or categories. The problem solved by the network is then a classification problem.

A classification problem can also be applied to textual data, documents, Internet pages, audio data.

The problem to be solved by the neural network can be very diverse in nature. It may be a problem of classifying elements into different categories in order to learn to recognize these elements.

The problem can also consist of the approximation of an unknown function or of an accelerated modeling of a known but complex function to calculate or even a problem of regression or prediction.

The present invention is more particularly applicable to a classification problem.

We consider a network called the source network, of the type described in section figure 1 , which is pre-trained on one or more training data sets to solve a source classification problem.

For example, the source problem consists in identifying, in a scene captured via one or more image(s), the presence of cars. In other words, this source problem consists in classifying the objects present in the scene according to different categories, one of them being the car category.

For this, one or more databases of training images are used which are used to train the network, in the manner described above, so as to optimize the synaptic weights of all the layers so that the network learns to classify objects corresponding to cars.

The picture 2 schematizes a second neural network, called the target network, which aims to learn how to solve a target classification problem, which is generally more specific than the source classification problem. For example, the target problem is to classify cars according to their type or brand.

A general goal of transfer-based machine learning is to exploit the optimized source network of the figure 1 to exploit its knowledge in order to solve the target problem and this, without carrying out a complete training on the same quantity of training data as for the source network.

For this, one possible method is to replace the last layer of the source network with an SM output layer specific to the target problem. In other words, the output layer SM of the target network obtained contains as many output neurons a1, a2, a3 as there are categories associated with the target problem (for example corresponding to the different categories of cars that it is desired to classify).

Then, we carry out a partial training of this network to update only the weights w1, w2, w3 of the last layer which are the most specific to the targeted problem. This partial training is carried out from new training data using a back-propagation algorithm in the same way as for the source network of the figure 1 , except that only the weights of the last layer are updated.

One objective of the invention is to carry out the adaptation of the source network into the target network without carrying out the partial learning, that is to say without using a complete back-propagation algorithm.

The figure 4 illustrates, on a flowchart, the steps for implementing an automatic learning transfer method according to a first embodiment of the invention.

The first step 401 consists, as explained above for the figure 1 , to carry out complete learning or training of the source neural network R _S from a set of source learning data DA _S with the aim of solving a source classification problem.

The second step 402 consists in replacing the last layer of the source network R _S by a layer specific to the target problem, in order to generate an architecture of a target network R _c The target network R _c differs from the source network R _S only by its last layer. Advantageously, the last layer of the target network is based on the SoftMax function as shown in picture 3 .

The picture 3 represents, in a non-limiting example, the last two layers of a target network R _c , that is to say the final SoftMax layer and the last hidden layer C _L . In this example, the SoftMax layer comprises three neurons a1,a2,a3 associated with three categories of the target classification problem. Each of these three neurons is linked to a vector of synaptic weights w ₁ ,w ₂ ,w ₃ .

The third step 403 takes as input a new set of training data intended to be used to solve the target classification problem. Typically this new data set is reduced in size compared to the first source training data set. For example, it concerns data acquired in an industrial context to solve a specific problem, the acquisition of this data being expensive and therefore the quantity of data available being limited. On the contrary, the source training data corresponds to a more general source problem and is therefore available in greater quantity. For example, these are publicly available databases such as ImageNet databases or other public databases for training neural networks for classification.

The third step 403 consists in propagating the target learning data from the input of the target network up to the last hidden layer C _L then in extracting the representations of these data from this last hidden layer. On the example of the picture 3 , the extracted representations correspond to the features x1,x2,x3,x4,x5.

From these representations, in a fourth step 404, the updated values of the synaptic weights w ₁ ,w ₂ ,w ₃ of the last layer are calculated as being proportional to the normalized centered average of the representations x1 ,x2,x3, x4,x5 and belonging to each respective class a1,a2,a3.

More precisely, the target training data used is labeled according to the category to which it belongs. These labels are propagated with the data until the last hidden layer.

Then, for each class associated with a label, the averages m _l of the representations on the set of learning data corresponding to this label are calculated.

, avec k le nombre de classes.The centered means are then calculated using the following formula: $\widetilde{m_I}=m_I-\frac{1}{k}{\displaystyle {\sum }_{I=1}^k{m}_I}$

, with k the number of classes.

avec Q_l un facteur de normalisation qui est, par exemple, égal ou proportionnel à $Q_l=\frac{1}{\Vert \widetilde{m_l}\Vert }$

.We finally calculate the weights of the last layer using the formula: $w_{\mathrm{you}}=Q_I\cdot \widetilde{m_I}$

with Q _l a normalization factor which is, for example, equal or proportional to $Q_I=\frac{1}{\Vert \widetilde{m_I}\Vert }$

.

This approximation of the calculation of the synaptic weights makes it possible to avoid having recourse to a backpropagation algorithm based on a gradient descent which has a significant cost in terms of the number of operations.

The proposed approximation is in particular valid for a final layer based on the SoftMax function and for odd activation functions (eg hyperbolic tangent, sine or linear) for the last hidden layer of the network.

, un vecteur dont les composantes correspondent à chaque classe du problème de classification.We recall that the Softmax function is defined by:
$f{\left(\mathrm{HAS}\right)}_I=\frac{e^{{\mathrm{has}}_I}}{{\displaystyle {\sum }_{k=1}^K{e}^{\mathrm{has}}k}}$

, a vector whose components correspond to each class of the classification problem.

This approximation is obtained by analyzing the theoretical behavior of the Softmax layer for classifier networks.

The figure 5 describes a second embodiment of the invention which aims to select the most suitable neural network for the target problem to be solved from among several pre-learned networks for a source problem.

Indeed, when several training databases are available, a problem to solve consists in determining which pre-trained network for which training database is the most suitable to solve the target problem.

In this second embodiment, the source neural network is trained to solve the source problem for several sets of learning data DA _S1 ,..., DA _SN . In a variant, several source neural networks having different architectures are trained for different sets of training data DA _S1 ,..., DA _SN .

At the end of the learning step 401, several pre-learned source neural networks R _S1 ,..., R _SN are therefore obtained, each corresponding to a set of learning data, for example to a database d public images in the case of an image classification problem, and/or a particular architecture,

Step 402 is applied to each pre-trained network by replacing the last layer with a layer specific to the target problem. Steps 403 and 404 of the method described in figure 4 to each network to obtain several target networks R _c1 ,..., R _cN from the same set of target training data DA _C .

In a final step 405, the classification performances of all the target networks R _c1 ,..., R _cN are compared and the target network R _ci which presents the best performances for solving the target problem is retained.

Step 405 is, for example, carried out by using labeled test data and by executing an inference phase of each target network for this test data.

At the end of the inference phase, the SoftMax layer of the target network calculates a probability or a classification score (as illustrated in picture 3 ).

The best performing target network is the one that provides the most reliable classification scores for the same test data.

Thus, this second embodiment makes it possible to select the pre-learned network best suited to the target problem without the need to carry out complete learning of the last layer of each pre-learned network.

The figure 6 describes a third embodiment of the invention, the objective of which is to select, among several sets of target training data available, the one which best makes it possible to improve the training of the pre-learned network in order to solve the target problem.

In this third embodiment, the target problem is identical to the source problem and therefore the last layer of the target network contains as many neurons as the last layer of the source network (unlike the previous embodiments).

An objective of this third embodiment is to determine which target training data set is most likely to modify the pre-trained model of the source network. In order to avoid redoing a complete training of the source network for each new set of target training data available, the target network is used to determine a criterion for selecting the most relevant target training data.

In this third embodiment, step 401 is identical to that described in figure 4 and is run for a single source training data set to produce a pre-trained source network model R _S .

Step 402 consists of replacing the last layer of the source network with a SoftMax layer but this new layer comprises as many neurons as the last layer of the source network so as to address the same source problem.

Steps 403,404 are executed for several sets of target training data DA _c1 , .., DA _cN . This scenario corresponds to a situation where new target training data is acquired over time and is available to update the source network. For example, in the case of a autonomous car, it can constantly acquire new images of its environment which can constitute new learning data.

At the end of step 404, several target networks R _c1 ,..., R _cN are obtained corresponding to each set of target training data.

A final step 605 is then applied to select the target network, and therefore the associated target training data set which makes it possible to deviate the most from the initially pre-learned model to generate the source network Rs. In other words, to From test data, an inference phase is executed for each target network and the classification scores determined by the SoftMax layer are compared.

We then retain the target network that gives the worst scores. Indeed, a bad classification score reflects the fact that the new target training data used is very far from the source training data DAs used to generate the source network model Rs.

Conversely, if the target network makes it possible to obtain a good classification score while the calculation of the weights of the last layer is based on a normalized centered mean criterion, then this means that the data used to update these weights are not very new compared to the source data.

By selecting the target training data DAc _i corresponding to the target network with the least performance in terms of classification, this means that these training data are really new with respect to the source training data DA _S .

Thus, it is useful to carry out a complete new training of the source network Rs from these new target training data DAc _i to update the model of the source network. We then loop back to step 401 to update the model of the source network Rs. The new training of the source network Rs can also be carried out by taking the source learning database DA _s and completing it with the new data learning target DAc _i .

An advantage of this third embodiment is that it avoids carrying out a complete new training of the source network for each new training data set available when these new data are regularly available over time. The invention makes it possible to define a criterion to evaluate the novelty of the new data available in order to select only this data to carry out a complete new learning of the source network.

The figure 7 represents a comparative example of performances obtained for a conventional learning method by back-propagation and for a method according to the invention.

The diagram of the figure 7 represents the classification performance as a function of the quantity of data tested over time respectively for the two aforementioned methods, during learning and during a test phase.

The curves 701,711 correspond to the performances obtained during learning respectively with a back-propagation method (701) and with a method according to the invention (711).

The curves 702,712 correspond to the performances obtained during an inference phase with test data for a network learned respectively with a back-propagation method (702) and with a method according to the invention (712).

The results of the figure 7 are obtained with a network whose last hidden layer comprises an activation function of the hyperbolic tangent type.

The curves of the diagram of the figure 7 show that the invention makes it possible to obtain classification performances similar to those obtained with conventional learning.

The invention can be implemented as a computer program comprising instructions for its execution. The computer program may be recorded on a processor-readable recording medium.

Reference to a computer program which, when executed, performs any of the functions previously described, is not limited to an application program running on a single host computer. Rather, the terms computer program and software are used herein in a general sense to refer to any type of computer code (e.g., application software, firmware, microcode, or other form of computer instruction) which can be used to program one or more processors for implement aspects of the techniques described herein. The computing means or resources can in particular be distributed (“ Cloud computing ”), possibly using peer-to-peer technologies. The software code may be executed on any suitable processor (e.g., a microprocessor) or processor core or set of processors, whether provided in a single computing device or distributed among multiple computing devices (e.g. example as possibly accessible in the environment of the device). The executable code of each program allowing the programmable device to implement the processes according to the invention can be stored, for example, in the hard disk or in ROM. In general, the program or programs can be loaded into one of the storage means of the device before being executed. The central unit can control and direct the execution of the instructions or portions of software code of the program or programs according to the invention, instructions which are stored in the hard disk or in the ROM or else in the other aforementioned storage elements.

The target neural network can be implemented on a computing device based, for example, on an embedded processor. The processor may be a generic processor, a specific processor, an application-specific integrated circuit (also known as an ASIC for "Application-Specific Integrated Circuit") or an array of field-programmable gates (also known as the English name of FPGA for “Field-Programmable Gate Array”). The computing device may use one or more dedicated electronic circuits or a general purpose circuit. The technique of the invention can be implemented on a reprogrammable calculation machine (a processor or a microcontroller for example) executing a program comprising a sequence of instructions, or on a dedicated calculation machine (for example a set of gates such as an FPGA or an ASIC, or any other hardware module).

The invention applies in many fields in which there is a need to solve a classification problem.

For example, the invention applies to the analysis of image or video data or more generally multimedia to solve a classification problem. In this case, the network input data is multimedia data.

The invention can be implemented in embedded devices, comprising several silicon layers respectively dedicated to the acquisition of input data (for example image, or video) and to calculations relating to the neural network.

The invention applies to any problem of classifying objects among data. For example, it may involve classifying images or audio sequences or textual data according to their content.

Claims (12)

A computer-implemented method of machine transfer learning comprising the steps of: - Training (401) an artificial neural network (R _S ) to solve a first source classification problem, from at least a first set of source learning data (DA _S ), - updating (404) the synaptic weights of the last layer of the trained artificial neural network, from a second set of target learning data (DA _c ), to solve a second target classification problem, - For each neuron of the last layer of the artificial neural network associated with a class, said synaptic weights being proportional to the normalized centered mean of the representations of the second set of target learning data (DA _c ) extracted from the last hidden layer of the artificial neural network and belonging to said class. Method according to claim 1 further comprising a step of replacing (402) the last layer of the trained artificial neural network with a layer specific to the second target classification problem. A method according to claim 2 wherein the specific layer is a layer based on the softmax function. A method according to any preceding claim wherein, to update the synaptic weights of the last layer of the trained artificial neural network, the target training data (DA _c ) of the second set is propagated (403) from the entry of said network up to its last hidden layer so as to be able to extract said representations therefrom. Method according to any one of the preceding claims, in which the artificial neural network is trained for several sets of source training data (DA _S1 ,..., DA _SN ), in order to produce several trained neural networks (R _S1 ,..., R _SN ), the updating of the synaptic weights is performed for each trained neural network and the method further comprises a step (405) of selecting the trained neural network ( R _ci ) which has the best performance in solving the second classification target problem. A method according to claim 5 wherein the second target training data set (DA _c ) is smaller in size than a source training data set. Method according to any one of Claims 1 to 4, in which the updating of the synaptic weights is carried out for several sets of target training data (DA _C1 ,..., DA _CN ), the second target problem is identical to the first source problem and the method further comprises a step of selecting (605) the target training data set which results in the worst performance in solving the second classification target problem. Method according to claim 7 further comprising a new step of training the artificial neural network from the selected target training data set or from a combination of the selected target training data set and the source training dataset. Method according to any one of the preceding claims, in which the last hidden layer of the artificial neural network comprises an odd activation function, for example a hyperbolic tangent function. Computing device for implementing an artificial neural network which has learned to solve a classification target problem using a learning transfer method according to any one of the preceding claims. Computer program downloadable from a communication network and/or stored on a medium readable by a processor, the program comprising instructions for the execution of the method according to one any of claims 1 to 9, when the program is executed by a processor. A processor-readable recording medium on which is recorded a program comprising instructions for the execution of the method according to any one of claims 1 to 9, when the program is executed by a processor.

Images